Building a browser-based ColorGPT with Jev
The author trained a 2.7M-parameter model that runs in the browser and outputs RGB distributions from a prompt, citing Jev's inference speed as a sign that small, fast, customizable models have a future.
jev reminds us how good speed can feel. i trained a colorgpt a while back, a 2.7m-parameter nanogpt-style transformer conditioned on frozen minilm embeddings and trained on the 2010 xkcd color survey. it runs in the browser and outputs only distributions over rgb values given a prompt. training took ~3 minutes on an rtx 5070, and the results aren't bad. it made me hopeful for a future of models in all shapes and sizes: fast, cheap, and easy to customize.
This post quotes another post: x.com/mattdesl/status/2100899669802963060